Rosenblatt Securities analysts “double downgraded” Arista Networks (NYSE:ANET) in April 2024 from a buy with stock price target of $330, to a sell with price target of $210, on rising competition from Nvidia (NASDAQ:NVDA). In the wake of this, there have been some highly unqualified opinions on what’s happening, and some atrocious “reporting” that is highly misleading for investors trying to dissect the business models.
Let’s clear the air with some Chip Stock Investor “radical moderateness,” because for most investors, we strongly believe there’s little value in swinging either too bullish or too bearish. Disclaimer: We are long-time shareholders of both Arista and Nvidia.
Arista and Nvidia build data centers, but Nvidia doesn’t like ethernet???
It would seem even the tech community, much less the financial community, knows little about what InfiniBand is, and how it competes (sort of) with ethernet.
Jim Keller, CPU design genius and CEO of TensTorrent, recently tweeted that Nvidia could have saved itself and its users a lot of money if only they had just used Ethernet instead of Infiniband to link the new Blackwell GPUS. This would have promoted more cohesion for customers that want to use software and or hardware from other companies. Which for obvious reasons, Nvidia doesn’t necessarily want. They are running a proprietary system for a reason. Hence the monstrous billions in data center revenue of the last several quarters.
Mr. Keller is a legend in the chip design community. But he’s also the CEO and co-founder of some of his own (privately held) businesses these days. A CEO role is a very different one from the executive chip designer roles he’s played in the past. Bear that in mind when posts like this come out.
For the record, Nvidia CEO Jensen Huang has now affirmed many times that “he loves ethernet” (more on that in a bit). Nvidia is putting its money where Jensen’s mouth is: Ethernet. And that’s what has some investors suddenly spooked.
How Arista (and Nvidia) actually make money
Arista and Nvidia are both in the business of engineering data centers, especially for tech giants that operate a cloud (Amazon AWS, Microsoft Azure, Google Cloud) or some sort of “consumer internet” product (Google Search, Meta Facebook/Instagram/WhatsApp).
Arista and Nvidia actually monetize this work of designing data centers, not by actually constructing them, but via the sale of hardware and software that gets designed into the ultimate computing facility. Just the infrastructure itself is worth ~$250 billion a year in global annual spend, a figure that’s now quickly rising as purpose-built data centers for generative AI training are being added to the traditional on-premises enterprise computing and public cloud markets.
Here’s how Arista and Nvidia each make money:
- For Arista, its data center switches, routers, other networking equipment, and operating software including cybersecurity services. In this case, an Ethernet-based switch connects multiple servers together to create a network; and an Ethernet router connects multiple switches and the network of servers they form, to make an even larger network. These are the devices used to create these data center “supercomputers” that big tech, government organizations, and the research community utilize today. Arista’s 2023 revenue was $5.86 billion, with a GAAP operating profit margin of 38.5%. From a product perspective, data center and cloud network hardware was 60%-65% of 2023 revenue, 10%-15% campus (on-premises) cognitive network hardware, and 20%-25% software and services. https://s21.q4cdn.com/861911615/files/doc_financials/2023/q4/irdeck_q4-23highlights_final4.pdf pg 19
- Nvidia makes money from its data center segment, famously from its GPU-based “AI” systems (or more broadly, accelerated computing systems), InfiniBand networking, Spectrum Ethernet networking, and various AI software libraries and AI management tools. Nvidia’s fiscal 2024 (mostly corresponding to calendar year 2023, since its last fiscal year ended in January 2024) was $60.9 billion, with a GAAP operating profit margin of 54.1%. $47.5 billion, or 78% of total sales, were lumped into Nvidia’s “Data Center” segment. https://s201.q4cdn.com/141608511/files/doc_financials/2024/Q4FY24/Rev_by_Mkt_Qtrly_Trend_Q424.pdf
Here’s what each company’s high-level financials looked like last year:
Enter Nvidia Spectrum Ethernet. Spectrum-X Ethernet hardware is not new, this is just the latest (and admittedly, by far the best to date) generation of Nvidia’s Ethernet networking gear. Spectrum too was acquired, along with InfiniBand products, from Mellanox back in 2020. We seriously doubt most tech analysts think Spectrum is new. But in case this is indeed a revelation, here’s a brief history.
Ethernet vs. InfiniBand
From an investor perspective, usually when we see Ethernet or InfiniBand or any other tech hardware for that matter, what we see is a product that gets sold and generates revenue. But behind every piece of hardware is intellectual property (IP), patents, and protocols that dictate how the hardware operates.
Sometimes IP and patents are proprietary, like Intel’s x86 processors (which by court order decades ago, AMD also licenses and sells: https://youtu.be/WGykUNea41Y https://chipstockinvestor.com/qualcomm-vs-apple-will-history-be-changed-as-two-new-chips-target-the-laptop-market-giants/ ). But sometimes hardware IP is open and available for all to use, if they want to dump the resources into R&D.
This is where Ethernet and InfiniBand come in, two sets of data networking open standards (remember the three main components of any computing system: Logic (computing), networking, and memory storage).
Ethernet and InfiniBand networking standards were codified decades ago, and a company can incorporate those standards and protocols into their hardware product and/or software.
Here’s a high-level overview of each, for the investor community.
Ethernet
Ethernet is prolific. Original standards were released back in 1980 for local area network (LAN) connections. Massive amounts of R&D have gone into Ethernet, making it almost a de facto option whenever hardware needs to be physically connected. If your PC is hardwired to your internet modem or router, or printer connected to PC, you’re likely looking at an Ethernet cable and ports they plug into.
Ethernet is also extensively used in business and high-performance compute (HPC). When you see those racks of servers (big powerful computers that run the internet, cloud services, and on-premises business applications only accessible via a private network), the bundle of cables coming out the back of those pull-out racks are likely ethernet too.
Arista has built is now sizable and still-growing data center empire (both for cloud, and private networks) on open architecture standards. Ethernet is the networking standard that Arista switches and routers utilize.
In 2023, Arista and a long list of others (including one of its top networking chip suppliers Broadcom) banded together to form the Ultra Ethernet Consortium to advance Ethernet technology standards for a new era of AI and HPC in the data center. https://ultraethernet.org/
InfiniBand
InfiniBand is the most commonly used networking interconnect type used in high-performance computing (HPC) in data center supercomputers. It was founded in 1999 via merger of two other HPC networking standards, to create the InfiniBand Trade Association, with initial standards released in 2000. https://www.infinibandta.org/about-the-ibta/
The dot-com bubble bust of the early 2000s banished InfiniBand to a type of IT purgatory for a time. Mellanox was founded in 1999, but various InfiniBand vendors started to head for the exit in favor of other networking standards products.
Intel closed down its InfiniBand switches in 2002, and later acquired the InfiniBand business from Qlogic in 2012 for just $125 million. Intel had also been toying around with another HPC networking standard, Omni-Path, which it discontinued in 2019. In 2020, it spun out the skeleton remnants of this business as Cornelis Networks, part of the Intel Capital investment portfolio. https://www.cnet.com/tech/tech-industry/intel-cancels-infiniband-products/ https://www.techpowerup.com/260020/intel-discontinues-omni-path-enabled-xeon-processors https://www.intel.com/content/www/us/en/customer-spotlight/stories/hlrn-is-meeting-diverse-needs-study.html
What was left of Qlogic after the Intel sale was acquired by network processor company Cavium in 2016, and Cavium itself was acquired by Marvell Technology Group in 2018. (Wow, that was seven years ago already??? Nick’s write-up on that: https://www.fool.com/investing/2017/12/02/the-marvell-technologycavium-deal-keeps-the-ball-r.aspx) https://www.reuters.com/article/idUSTRE80M173/ https://www.businesswire.com/news/home/20160615006497/en/Cavium-to-Acquire-QLogic https://www.marvell.com/company/newsroom/marvell-technology-completes-acquisition-of-cavium.html
Intel and Marvell remain part of the InfiniBand Trade Association. Intel, along with IBM, Hewlett-Packard Enterprise, and of course Nvidia acting as the “Steering Committee Members.” https://www.infinibandta.org/member-listing/
Meanwhile, Microsoft also withdrew InfiniBand development support in 2002. One of the last major supporters of the HPC networking standards was SunMicrosystems, which was acquired by Oracle in 2010. Not surprisingly, Oracle was an early adopter of Nvidia’s AI hardware in 2022, putting InfiniBand to liberal use in building out Oracle Cloud with all those Nvidia GPUs. (Our write-up on that: https://www.fool.com/investing/2022/10/28/excited-about-the-work-oracle-cloud-is-doing-not-s/ ) https://www.zdnet.com/home-and-office/networking/microsoft-backs-away-from-infiniband/ https://www.oracle.com/corporate/pressrelease/oracle-buys-sun-042009.html
This meant there was really only one major InfiniBand hardware R&D and product company in 2010 when Mellanox merged with another small InfiniBand company called Voltaire, with Intel sealing the fate with that purchase of the Qlogic business two years later. Thus, when Nvidia announced it was acquiring this little networking business called Mellanox in 2019 (which finally received regulatory approval and was completed in early 2020), not many were paying much attention – not in the tech world, nor in the financial world. (Our original write-ups in 2019 and 2020 on the topic: https://www.fool.com/investing/2019/10/12/where-will-nvidia-be-in-5-years.aspx https://www.fool.com/investing/2020/02/18/with-nvidia-revenue-surge-some-investors-were.aspx https://www.fool.com/investing/2020/04/19/nvidia-mellanox-finally-okd-by-chinese-regulators.aspx https://www.fool.com/investing/2020/05/26/nvidia-data-center-business-boost-mellanox.aspx )
You can see our video from late summer 2023 where we further explained how the Mellanox purchase could go down in history as one o the greatest buys ever: https://www.youtube.com/watch?v=upv4IShQFZM
Why InfiniBand for AI?
Theoretically InfiniBand is an open set of standards, incorporated into the Linux open-source software list for data centers. But in reality, this networking technology is now controlled by Nvidia, and it’s making some serious dough. We estimate that as much as one-third of Nvidia’s “data center” revenue could be coming from InfiniBand. https://network.nvidia.com/pdf/whitepapers/WP_2007_IB_Software_and_Protocols.pdf
But why is it better for AI, and in particular AI training, than Ethernet? We’ll let Jensen himself explain:
That’s a powerful illustration. Ethernet was originally designed for data to get from point A to point B as quickly as possible, and to average all the data in a network so that data all gets to point B quickly. But if Ethernet were used to network together GPUs for AI training, it could leave some GPUs “finishing their homework early,” and then sitting idle while others gather up and process their data later. The InfiniBand breakthrough was monitoring both point A and point B in a network, so that data is constantly flowing to all parts of the network to create one cohesive data center supercomputer cluster working in orchestration.
Put even more simply, Ethernet is fantastic for non-AI data center HPC, and even for AI inference (after the AI has been trained and is now being used). Arista has been making good money off of its AI efforts for new use cases that are being demanded from data center customers.
But InfiniBand is the best bet at this point for AI training networking. Here’s one of Nvidia’s YouTube videos on how all of this works: https://www.youtube.com/watch?v=nKqfi3q4S5I
So Nvidia must really hate Ethernet, right?
Wrong! Again, as alluded to earlier, Jensen has publicly said multiple times “he loves Ethernet.” Along with InfiniBand products, Mellanox also came with Spectrum Ethernet products. We outlined the full data center product list a couple months back in a video here: https://youtu.be/Rmz3GpexOrs
And here’s that Nvidia data center product list one more time:
Data source: Nvidia. https://www.nvidia.com/en-us/data-center/technologies/hopper-architecture/ https://www.nvidia.com/en-us/data-center/grace-cpu/ https://www.nvidia.com/en-us/networking/products/data-processing-unit/ https://www.nvidia.com/en-us/data-center/nvlink-c2c/ http://nvidianews.nvidia.com/news/nvidia-completes-acquisition-of-mellanox-creating-major-force-driving-next-gen-data-centers https://www.nvidia.com/en-us/networking/products/infiniband/
Spectrum Ethernet is not new, then, but the latest generation – Spectrum X – seems to have Rosenblatt suddenly worried that Arista’s Ethernet-based business could come under threat from Nvidia. It’s certainly a valid concern. But Spectrum X was actually announced in spring 2023. Again, this is far from new information. https://nvidianews.nvidia.com/news/nvidia-launches-accelerated-ethernet-platform-for-hyperscale-generative-ai
Per Jensen in some comments in March 2024:
So what’s really going on? And is Arista toast?
The important bits of the quote above are that yes, Arista (and its key chip design partners like Broadcom) have new competition from Nvidia. But even Jensen acknowledges that Ethernet has the ability to strike back against the InfiniBand AI training use case within the next few years, and can be further developed to make HPC workloads even better.
That’s why we recently covered Broadcom’s own AI event the same week as Nvidia’s GTC in March 2024. Broadcom is really leaning into its Ethernet capabilities, especially for its custom chip design business for big data center operators. https://www.broadcom.com/company/news/product-releases/61571 https://youtu.be/0ST3Ak52DA4
What is Rosenblatt getting at?
We see two possibilities here.
- The analyst team at Rosenblatt weren’t aware of the two different Nvidia networking products until just recently (a small, but not 0%, possibility), but we can’t say for sure, we don’t diligently follow Rosenblatt or any other analyst; or
- Rosenblatt is acting as an important “agent of chaos” market mechanism that keeps the stock valuation in check
The more likely second scenario is totally fine, although we have misgivings about the “double downgrade.” If you knew about Spectrum Ethernet all along, but still maintained near-term ANET stock targets at $330 per share (which implied a TTM PE ratio of ~50x), and suddenly decided the business is worth over a third less than that, we have questions. Not much has really changed.
This is one reason we don’t do short-term stock price targets.
Ultimately, an important market feature was introduced. Risks must be accounted for to keep valuations tame. Arista is a superbly run business (in our opinion), but there always need to be limits to the premium we place on quality. This is why we were holding, but not adding to, our position in the stock when we reviewed it in February. Management already signaled slower growth in 2024.
Investors still warming up to long-term Arista ownership should only be using a small dollar-cost average plan, as shares have only “crashed” back to where they were two months ago after the analyst downgrade. https://www.youtube.com/watch?v=cxQxhmh6c_s
Arista Networks stock was cheap for much of the last couple of years, but that “cheapness” is gone. Rosenblatt helped get us back closer to value, even though the method in arriving there might have been a bit odd.